This action might not be possible to undo. Are you sure you want to continue?

**Patrick Guillaume, Department of Mechanical Engineering,Vrije Universiteit Brussel, Pleinlaan
**

2, B-1050 Brussel, Belgium.

Keywords: Vibration, Estimation, Frequency domain, Modal analysis, Modal parameters, Natural

frequency, Damping, Mode shapes, Transfer function, SISO, MIMO, Mechanical systems, SDOF,

MDOF.

Contents

1. Introduction

2. The “Modal” Model

2.1. Single Degree of Freedom

2.2. Multiple Degree of Freedom

2.2.1. Mode Shapes and Operating Deflection Shapes

2.2.2. Observability and Controllability of Modes

3. Frequency-Domain Identification of Modes

3.1. Least-Squares Estimation

3.1.1. Common-Denominator Model

3.1.2. Linearity in the Parameters

3.1.3. Normal Equations

3.1.4. Reduced Normal Equations

3.1.5. Stabilization Chart

3.2. Maximum Likelihood Estimation

3.2.1. Gauss-Newton Optimization

3.2.2. Confidence Intervals

4. Application

5. Conclusion

Glossary

DOF: Degree of freedom.

FRF: Frequency response function.

GTLS: Generalized total least squares.

IQML: Iterative quadratic maximum likelihood.

IRF: Impulse response function.

LS: Least squares.

LSCE: Least squares complex exponential.

MDOF: Multiple-degree-of-freedom.

ML: Maximum likelihood.

MIMO: Multiple-input-multiple-output

SDOF: Single-degree-of-freedom.

SISO: Single-input-single-output.

TLS: Total least squares.

Summary

In this contribution the applicability of frequency-domain estimators in the field of modal analysis

will be illustrated. The basics of vibration and modal analysis are briefly summarized. In modal

analysis, mechanical systems with a few inputs and hundreds of outputs have to be identified. This

requires adapted frequency-domain estimators designed to handle large amount of data in a

reasonable amount of time. A practical example will be given and finally the conclusions will be

drawn.

1. Introduction

It is well known that (mechanical) structures can resonate, i.e. that small forces can result in

important deformation, and possibly, damage can be induced in the structure.

Figure 1: Tacoma Narrows Bridge Disaster.

The Tacoma Narrows bridge disaster (Figure 1) is a typical example of this. On November 7, 1940,

the Tacoma Narrows suspension bridge collapsed due to wind-induced vibration (i.e. flutter).

Situated on the Tacoma Narrows in Puget Sound, near the city of Tacoma, Washington, the bridge

had only been open for traffic a few months.

Wings of airplanes can be subjected to similar flutter phenomena during flight. Before an airplane is

released, flight flutter tests have to be performed to detect possible onset of flutter. The classical

flight flutter testing approach is to expand the flight envelope of a airplane by performing a

vibration test at constant flight conditions, curve-fit the data to estimate the resonance frequencies

and damping ratios, and then to plot these frequencies and damping estimates against flight speed or

Mach number. The damping values are then extrapolated in order to determine whether it is save to

proceed to the next flight test point. Flutter will occur when one of the damping values tends to

become negative. Before starting the flight tests, ground vibration tests as well as numerical

simulations and wind tunnel tests (see Figure 2) are used to get some prior insight into the problem.

(a) (b)

Figure 2: Wind tunnel tests on a scaled model of (a) a Cessna and (b) an Airbus A380.

The majority of structures can be made to resonate, i.e. to vibrate with excessive oscillatory motion.

Resonant vibration is mainly caused by an interaction between the inertial and elastic properties of

the materials within a structure. Resonance is often the cause of, or at least a contributing factor to

many of the vibration and noise related problems that occur in structures and operating machinery.

To better understand any structural vibration problem, the resonant frequencies of a structure need

to be identified and quantified. Today, modal analysis has become a widespread means of finding

the modes of vibration of a machine or structure (Figure 3). In every development of a new or

improved mechanical product, structural dynamics testing on product prototypes is used to assess its

real dynamic behavior.

Figure 3: Modal analysis of a car body.

2. The “Modal” Model

Modes are inherent properties of a structure, and are determined by the material properties (mass,

damping, and stiffness), and boundary conditions of the structure. Each mode is defined by a natural

(modal or resonant) frequency, modal damping, and a mode shape (i.e. the so-called “modal

parameters”). If either the material properties or the boundary conditions of a structure change, its

modes will change. For instance, if mass is added to a structure, it will vibrate differently. To

understand this, we will make use of the concept of single and multiple-degree-of-freedom systems.

2.1 Single Degree of Freedom

A single-degree-of-freedom (SDOF) system (see Figure 4 where the mass m can only move along

the vertical x-axis) is described by the following equation

) ( ) ( ) ( ) ( t f t kx t x c t x m · + + (1)

with m the mass, c the damping coefficient, and k the stiffness. This equation states that the sum of

all forces acting on the mass m should be equal to zero with ) (t f an externally applied force,

) (t x m − the inertial force, ) (t x c − the (viscous) damping force, and ) (t kx − the restoring force. The

variable ) (t x stands for the position of the mass m with respect to its equilibrium point, i.e. the

position of the mass when 0 ) ( ≡ t f . Transforming (1) to the Laplace domain (assuming zero initial

conditions) yields

) ( ) ( ) ( s F s X s Z · (2)

with ) (s Z the dynamic stiffness

k cs ms s Z + + ·

2

) ( (3)

The transfer function ) (s H between displacement and force, ) ( ) ( ) ( s F s H s X · , equals the inverse

of the dynamic stiffness

k cs ms

s H

+ +

·

2

1

) ( (4)

Figure 4: SDOF system.

The roots of the denominator of the transfer function, i.e. k cs ms s d + + ·

2

) ( , are the poles of the

system. In mechanical structures, the damping coefficient c is usually very small resulting in a

complex conjugate pole pair

d

ω σ λ i t − · (5)

with π ω 2

d d

f · the damped natural frequency,

π ω 2

n n

f · the (undamped) natural frequency where λ ω · · m k

n

, and

λ σ ω ζ · ·

n

m c 2 the damping ratio (

2

1 ζ − ·

n d

f f ).

If, for instance, a mass m ∆ is added to the original mass m of the structure, its natural frequency

decreases to ) ( m m k

n

∆ + · ω . If 0 · c , the system is not damped and the poles becomes purely

imaginary,

n

ω λ i t · .

The Frequency Response Function (FRF), denoted by ) (ω H , is obtain by replacing the Laplace

variable s in (4) by ω i resulting in

ω ω ω ω

ω

c m k k c m

H

i ) (

1

i

1

) (

2 2

+ −

·

+ + −

· (6)

Clearly, if 0 · c , then ) (ω H goes to infinity for m k

n

· →ω ω (see Figure 4).

Although very few practical structures could realistically be modeled by a single-degree-of-freedom

(SDOF) system, the properties of such a system are important because those of a more complex

multiple-degree-of-freedom (MDOF) system can always be represented as the linear superposition

of a number of SDOF characteristics (when the system is linear time-invariant).

2.2 Multiple Degree of Freedom

Multiple-degree-of-freedom (MDOF) systems are described by the following equation

) ( ) ( ) ( ) ( t t t t f Kx x C x M · + + (7)

In Figure 5, the different matrices are defined for a 2-DOF system with both DOF along the vertical

x-axis.

x

1

(t)

x

2

(t)

f

1

(t)

f

2

(t)

m

2

k

1

c

1

m

1

c

2

k

2

x

1

(t)

x

2

(t)

f

1

(t)

f

2

(t)

m

2

k

1

c

1

m

1

c

2

k

2

1

]

1

¸

·

2

1

0

0

m

m

M

1

]

1

¸

−

− +

·

2 2

2 2 1

k k

k k k

K

1

]

1

¸

−

− +

·

2 2

2 2 1

c c

c c c

C

¹

;

¹

¹

'

¹

·

) (

) (

) (

2

1

t f

t f

t f

¹

;

¹

¹

'

¹

·

) (

) (

) (

2

1

t x

t x

t x

Figure 5: 2-DOF system.

Transforming (7) to the Laplace domain (assuming zero initial conditions) yields

) ( ) ( ) s s s F X Z( · (8)

with ) (s Z the dynamic stiffness matrix

K C M Z + + · s s s

2

) ( (9)

The transfer function matrix ) (s H between displacement and force vectors, ) ( ) ( ) ( s s s F H X · ,

equals the inverse of the dynamic stiffness matrix

) (

) (

] [ ) (

1 2

s d

s

s s s

N

K C M H · + + ·

−

(10)

with the numerator polynomial matrix ) (s N given by

) ( ) (

2

K C M adj N + + · s s s (11)

and the common-denominator polynomial ) (s d , also known as the characteristic polynomial,

) det( ) (

2

K C M + + · s s s d (12)

When the damping is small, the roots of the characteristic polynomial ) (s d are complex conjugate

pole pairs,

m

λ and

∗

m

λ ,

m

N m , , 1 · , with

m

N the number of modes of the system. The transfer

function can be rewritten in a pole-residue form, i.e. the so-called “modal” model (assuming all

poles have multiplicity one)

∑

·

∗

∗

−

+

−

·

m

N

m m

m

m

m

s s

s

1

) (

λ λ

R R

H (13)

The residue matrices

m

R ,

m

N m , , 1 · , are defined by

) )( ( lim

m

s

m

s s

m

λ

λ

− ·

→

H R (14)

It can be shown that the matrix

m

R is of rank one meaning that

m

R can be decomposed as

¸ ]

) ( ) 2 ( ) 1 (

) (

) 2 (

) 1 (

m m m m

m m

m

m

T

m m m

N

N

ψ ψ ψ

ψ

ψ

ψ

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

'

¹

· · R (15)

with

m

a vector representing the “mode shape” of mode m. From equation (13), one concludes

that the transfer function matrix of a linear time-invariant MDOF system with

m

N DOFs is the sum

of

m

N SDOF transfer functions (“modal superposition”) and that the full transfer function matrix is

completely characterized by the modal parameters, i.e. the poles

m d m m

i

,

ω σ λ t − · and the mode

shape vectors

m

,

m

N m , , 1 · .

Taking the inverse Laplace transform of (13) gives the Impulse Response Function (IRF)

∑

·

∗

∗

+ ·

m

m m

N

m

t

m

t

m

e e t

1

) (

λ λ

R R h (16)

which consists of a sum of complex exponential functions.

2.2.1 Mode Shapes and Operating Deflection Shapes

At or near the natural frequency of a mode, the overall vibration shape (“operating deflection

shape”) of a structure will tend to be dominated by the mode shape of the resonance. Applying a

harmonic force at one of the DOFs, say r, with angular frequency corresponding to the damped

natural frequency of for instance the n-th mode, i.e.

n d

i s

,

ω · , results in a displacement vector that

is approximately equal to

) (

) (

) (

, , n d r

n

n n

n d

F

r

ω

σ

ψ

ω X ≈ (17)

The observed displacement at angular frequency

n d ,

ω is called an “operating deflection shapes” and

is approximately proportional to the mode shape vector of mode n, i.e.

n

. When

n

σ is small, the

proportionality is well satisfied. In reality, there will always be a (small) contribution of the other

modes resulting in

) (

) ( i

) (

) ( i

) ( ) (

) (

,

1 1

, n d r

N

m m n m

m m

N

n m

m m n m

m m

n

n n

n d

F

r r r

m m

ω

ω ω σ

ψ

ω ω σ

ψ

σ

ψ

ω

,

_

¸

¸

+ +

+

− +

+ ·

∑ ∑

·

∗ ∗

≠

·

X (18)

2.2.2 Observability and Controllability of Modes

Assuming, for example, that one force is applied in DOF 1 while the displacement is observed in

DOF 2. In that case, the multiple-input-multiple-output (MIMO) transfer function matrix simplifies

to the following single-input-single-output (SISO) transfer function

∑

·

∗

∗ ∗

−

⋅

+

−

⋅

·

m

N

m m

m m

m

m m

s s

s H

1

1 , 2

) 1 ( ) 2 ( ) 1 ( ) 2 (

) (

λ

ψ ψ

λ

ψ ψ

(19)

If 0 ) 1 ( ≠

n

ψ then mode n will only be “observed” in DOF 2 if 0 ) 2 ( ≠

n

ψ . If 0 ) 1 ( ·

n

ψ then it is

clear that the terms corresponding to mode n will not appear in the sum, i.e. mode n cannot be

excited (or “controlled”) by applying a force in DOF 1. The DOFs where a mode shape vector

equals zero are called nodal points or nodes. In practice, this means that the force actuator should

not be positioned in a nodal point of the modes of interest. To reduce the risk of missing modes, the

number of excitation points can be increased. The same is true for the response measurements. The

number of inputs (excitation points) is typically in the order of 1 to 10, while the number of outputs

(response measurements) can reach more than 1000 points when using optical measurement

equipment such as for instance a scanning laser Doppler vibrometer.

3. Frequency-Domain Identification of Modes

Typical for modal analysis is the very large number of outputs. This huge amount of data requires

dedicated algorithms that balance between accuracy and memory/computing needs. In Section 3.1

such a ‘dedicated’ frequency-domain least-squares estimator will be presented. Based on these

results, it is possible to implement more sophisticated identification methods (see 6.43.8.2

Estimation with Known Noise Model and 6.43.8.4 Estimation with Unknown Noise Model) such as

for instance the frequency-domain Maximum Likelihood (ML) estimator (Section 3.2).

3.1 Least Squares Estimation

3.1.1 Common-Denominator Model

The relationship between output o (

o

N o , , 1 · ) and input i (

i

N i , , 1 · ) is modeled in the

frequency domain by means of a common-denominator transfer function

) (

) (

) (

ˆ

ω

ω

ω

d

N

H

k

k

· (20)

for

i o

N N k , , 1 · (where i N o k

i

+ − · ) 1 ( ) and with

∑

·

Ω ·

n

j

kj j k

B N

0

) ( ) ( ω ω (21)

the numerator polynomial between output o and input i and

∑

·

Ω ·

n

j

j j

A d

0

) ( ) ( ω ω (22)

the common-denominator polynomial. The real-valued coefficients

j

A and

kj

B are the parameters

to be estimated. Several choices are possible for the polynomial basis functions ) (ω

j

Ω . For a

discrete-time domain model, the functions ) (ω

j

Ω are usually given by ) i exp( ) ( j T

s j

⋅ − · Ω ω ω

(with

s

T the sampling period) while for a continuous-time domain model

j

j

) i ( ) ( ω ω · Ω . The bad

numerical conditioning of the continuous-time domain approach can be improved by using for

instance orthogonal Forsythe polynomials (at the expense of an increase of the computation time).

3.1.2 Linearity in the Parameters

In modal analysis, measurements of Frequency Response Functions (FRFs) are commonly used (see

6.43.8.1 Measurements of Frequency response functions). Replacing the model ) (

ˆ

ω

k

H in (20) by

the measured FRFs ) (

f k

H ω for

f

N f , , 1 · gives, after multiplication with the denominator

polynomial,

0 ) ( ) ( ) (

0 0

≈ Ω − Ω

∑ ∑

· ·

n

j

n

j

j f k f j kj f j

A H B ω ω ω (23)

with

i o

N N k , , 1 · and

f

N f , , 1 · . Because the equations (23) are “linear-in-the-parameters”,

they can be reformulated as

0

0 0

0 0

0 0

2

1

2 2

1 1

≈

¹

¹

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

¹

¹

'

¹

⋅

1

1

1

1

]

1

¸

Y X

Y X

Y X

i o

i o i o

N N

N N N N

(24)

with

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

'

¹

·

kn

k

k

k

B

B

B

1

0

,

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

'

¹

·

n

A

A

A

1

0

(25)

1

1

1

]

1

¸

Ω Ω Ω

Ω Ω Ω

·

] ) ( , , ) ( , ) ( )[ (

] ) ( , , ) ( , ) ( )[ (

1 0

1 1 1 1 0 1

f f f f

N n N N N k

n k

k

W

W

ω ω ω ω

ω ω ω ω

X (26)

1

1

1

]

1

¸

Ω Ω Ω −

Ω Ω Ω −

·

] ) ( ) ( , , ) ( ) ( , ) ( ) ( )[ (

] ) ( ) ( , , ) ( ) ( , ) ( ) ( )[ (

1 0

1 1 1 1 1 1 1 0 1

f f f f f f f

N k N n N k N N k N N k

k n k k k

k

H H H W

H H H W

ω ω ω ω ω ω ω

ω ω ω ω ω ω ω

Y (27)

Note that every equation in (23) has been weighted with a frequency-dependent function ) (

f k

W ω .

The quality of the estimates can often be improved by using an adequate weighting function. The

(complex) Jacobian matrix J of this least-squares problem

1

1

1

1

]

1

¸

·

i o i o

N N N N

Y X

Y X

Y X

J

0 0

0 0

0 0

2 2

1 1

(28)

has N

f

N

o

N

i

rows and (n+1)(N

o

N

i

+1) columns (with N

f

>> n, where n is the order of the

polynomials,

m

N n 2 · ).

3.1.3 Normal Equations

Many estimators used in modal analysis form the normal equations explicitly, i.e. they compute

) Re( J J

H

explicitly. Note that the real part of J J

H

has to be taken because the coefficients are real.

Deriving the estimates directly from the Jacobian matrix leads to a better-conditioned problem.

However, forming the normal equations can result in a faster implementation, as will be the case

here too. The normal equations can be written as

0

0

0

2

1

1

2 1

2 2

1 1

≈

¹

¹

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

¹

¹

'

¹

⋅

1

1

1

1

1

]

1

¸

∑

·

T S S

S R

S R

i o

i o

N N

N N

k

k

T T

(29)

with ) Re(

k

H

k k

X X R · , ) Re(

k

H

k k

Y X S · , and ) Re(

k

H

k k

Y Y T · . The entries of these matrices equal

,

_

¸

¸

Ω Ω ⋅ ·

,

_

¸

¸

Ω Ω ⋅ − ·

,

_

¸

¸

Ω Ω ⋅ ·

∑

∑

∑

·

− −

·

− −

·

− −

f

f

f

N

f

f s f

H

r f k f k k

N

f

f s f

H

r f k f k k

N

f

f s f

H

r f k k

H W s r T

H W s r S

W s r R

1

1 1

2

1

1 1

2

1

1 1

2

) ( ) ( ) ( ) ( Re ) , (

) ( ) ( ) ( ) ( Re ) , (

) ( ) ( ) ( Re ) , (

ω ω ω ω

ω ω ω ω

ω ω ω

(30)

If a discrete time-domain model is used, i.e. ) i exp( ) ( j T

s f f j

⋅ − · Ω ω ω , and if the frequencies are

uniformly distributed (i.e. ω ω ∆ ⋅ · f

f

,

f

N f , , 1 · , with

s

NT π ω 2 · ∆ ), then, the above

summations can be rewritten as

,

_

¸

¸

⋅ ·

,

_

¸

¸

⋅ − ·

,

_

¸

¸

⋅ ·

∑

∑

∑

·

−

·

−

·

−

f

f

f

N

f

N f s r

f k f k k

N

f

N f s r

f k f k k

N

f

N f s r

f k k

e H W s r T

e H W s r S

e W s r R

1

) ( 2 i

2

1

) ( 2 i

2

1

) ( 2 i

2

) ( ) ( Re ) , (

) ( ) ( Re ) , (

) ( Re ) , (

π

π

π

ω ω

ω ω

ω

(31)

One can readily verify that the above matrices have a Toeplitz structure and that their entries can be

time-efficiently computed with the Fast Fourier Transform (FFT) algorithm.

3.1.4 Reduced Normal Equations

Although the number of rows of the normal matrix in (29) is much smaller than the number of rows

of the Jacobian matrix (28), its size is still quite huge (i.e. (n+1)(N

o

N

i

+1) rows and columns). As

we are mainly interested in a fast and stable method to construct a stabilization chart (see next

section), only the denominator coefficients (i.e. the poles) are in fact required. Elimination of the

numerator coefficients

S R ⋅ ⋅ − ·

−

k k k

1

(32)

yields

0 S R S T ≈ ⋅

1

]

1

¸

⋅ ⋅ −

∑

·

−

i o

N N

k

k k

T

k k

1

1

(33)

or 0 M ≈ ⋅ with

∑

·

−

⋅ ⋅ − ·

i o

N N

k

k k

T

k k

1

1

S R S T M . The size of the square matrix M is n+1, and thus

much smaller than the original normal equation (29). To remove the parameter redundancy of

transfer function model (20) (and to avoid the trivial solution with all coefficient equal to zero), a

constraint has to be imposed on the coefficients of the transfer functions. This can be done, for

instance, by imposing that one of the coefficients is equal to a non-zero constant value. Assume, for

instance, that the last coefficient of D is constrained to 1 (i.e. coefficient n+1). In that case, the

Least Squares (LS) estimate of D is given by

¹

;

¹

¹

'

¹ + ⋅ −

·

−

1

)} 1 , : 1 ( { )] : 1 , : 1 ( [

ˆ

1

LS

n n n n M M

(34)

Once

LS

ˆ is known, (32) can be used to derive all

LS

ˆ

coefficients. This approach is more time

efficient that solving (29) directly (approximately

2 2

i o

N N times faster). The mode shape vectors are

derived using (14) and (15).

3.1.5 Stabilization chart

In modal analysis, a stabilization chart is an important tool that is often used to assist the user in

separating physical poles from mathematical ones. A stabilization chart is obtained by repeating the

analysis for increasing model order n. For each model order, the poles are calculated from the

estimated denominator coefficients. The stable poles (i.e. the poles with a negative real part) are

then presented graphically in ascending model order in a so-called “stabilization chart” (see Figure

6). Estimated poles corresponding to physically relevant system modes tend to appear for each

estimation order at nearly identical locations, while the so-called mathematical poles, i.e. poles

resulting from the mathematical solution of the normal equations but meaningless with respect to

the physical interpretation, tend to jump around. These mathematical poles are mainly due to the

presence of noise on the measurements.

(a) (b)

Figure 6. Stabilization chart obtained with (a) a time-domain estimator (LSCE) and

(b) the frequency-domain least-squares estimator.

The LSCE estimator (Least Squares Complex Exponential) is probably the most frequently used

technique in industry. The LSCE estimator is a time-domain technique that makes use of impulse

response functions (16) to derive the modal parameters. In Figure 6 the stabilization chart of the

LSCE estimator is compared with the proposed frequency-domain least squares estimator. It turns

out that in many applications, the frequency-domain estimator is able to generate quite clear

stabilization compared to the LSCE approach.

3.2 Maximum Likelihood Estimation

3.2.1 Gauss-Newton Optimization

Assuming the FRFs to be uncorrelated, the (negative) log-likelihood function reduces to

∑∑

· ·

−

·

i o f N N

k

N

f f k

f k f k

H

H H

1 1

2

ML

)} ( var{

) ( ) , (

ˆ

) (

ω

ω ω

" (35)

The Maximum Likelihood (ML) estimate of

T T T

N N

T

i o

] , , , [

1

· is obtained by minimizing (35).

This can be done by means of a Gauss-Newton optimization algorithm, which takes advantage of

the quadratic form of the cost function (35). The Gauss-Newton iterations are given by

p p p

p p

H

p p p

H

p

r J J J

+ ·

− ·

+1

set (b)

for ) Re( ) Re( solve (a)

(36)

with ) (

p p

r r · ,

p

r

p

J ∂ ∂ · ) ( and

¹

¹

¹

¹

¹

¹

¹

;

¹

¹

¹

¹

¹

¹

¹

¹

'

¹

−

−

·

)} ( var{

) ( ) , (

ˆ

)} ( var{

) ( ) , (

ˆ

) (

1 11

1 11 1 11

f i o

f i o f i o

N N N

N N N N N N

H

H H

H

H H

ω

ω ω

ω

ω ω

r (37)

The Jacobian matrix

p

J has the same structure as the matrix J given in (28). Also here it is possible

to form the normal equations (i.e. ) Re(

p

H

p

J J and ) Re(

p

H

p

r J ) in a similar time-efficient way as

presented in Section 3.1. See 6.43.8.2 Estimation with Known Noise Model and 6.43.8.4 Estimation

with Unknown Noise Model for more information about frequency-domain ML identification.

3.2.2 Confidence Intervals

A good approximation of the covariance matrix of the ML estimate

ML

ˆ

is obtained by inverting the

Fisher information matrix (see 6.43.8. Frequency Domain System Identification)

1

ML

)] Re( 2 [ }

ˆ

cov{

−

∞ ∞

≈ J J

H

(38)

with

∞

J the Jacobian matrix evaluated in the last iteration step of the Gauss-Newton optimization.

As one is mainly interested in the uncertainty on the modal frequencies and damping ratios, only the

covariance matrix of the denominator coefficients is in fact required. Starting from (38), one can

show that this matrix is given by

1

1

1

ML

2 } ˆ cov{

−

·

−

1

]

1

¸

⋅ ⋅ − ≈

∑

i o

N N

k

k k

T

k k

S R S T (39)

with

k

R ,

k

S , and

k

T as defined in Section 3.1.3 but now applied to ) Re(

∞ ∞

J J

H

. Hence, it is not

necessary to invert the full matrix occurring in (38). From (39), it is possible to compute the

uncertainty on the modal frequencies and damping ratios. For flight flutter testing – but also for

applications such as vibration-based fault detection and operational modal analysis – the availability

of reliable estimates together with confidence intervals is important.

4. Application

In this section, modal analysis will be applied to a slat track, which is safety critical component of a

airplane. Slat tracks (see Figure 7) are located at the leading edge of an aircraft wing and make part

of a gliding mechanism that is used to enlarge the wing surface. The enlargement of the wing

surface is needed in order to increase the lift force at reduced velocity during landing and take off.

A slat track of an Airbus A320 airplane is considered here. The A320 airplane has 5 slats per wing.

The first slat (i.e. the inboard slat between fuselage and engine) contains 4 slat tracks. The other 4

slats have 2 slat tracks each. Safety critical components such as slat tracks are rigorously tested to

prove their ability to withstand all safety regulations. It is commonly accepted that the track should

outlive the plane by five times. Using computer simulations, it is possible to predict the lifetime of

a track using Finite Element (FE) models. To validate the dynamic behavior of these FE models,

experimentally obtained estimates of the modal parameters are required.

SLAT

SLAT TRACK

LEADING

EDGE

OF WING

SLAT

SLAT TRACK

LEADING

EDGE

OF WING

Figure 7: Mounted slat track on an Airbus A320 in extended position.

SHAKER

STINGER

FORCE SENSOR

SLAT TRACK

SHAKER

STINGER

FORCE SENSOR

SLAT TRACK

SHAKER

SCANNING LASER

DOPPLER VIBROMETER

SHAKER

SCANNING LASER

DOPPLER VIBROMETER

(a) (b) (c)

Figure 8: Measurement setup. (a) Excitation with a shaker. (b) Response measurements with a

scanning laser Doppler vibrometer. (c) Measurement points on the surface of the slat track.

In Figure 8(a) a shaker is used to excite the slat track with a multisine signal. A multisine is a

periodic signal consisting of a sum of sine waves at uniformly distributed frequencies (from 0 Hz to

8192 Hz in this case with a resolution of 1 Hz). A force sensor is attached on the slat track and is in

contact with the shaker through a stinger. The stinger has the characteristic of being stiff in only

one direction, i.e. that of the intended excitation. In this example, one surface of the track is

measured by means of a scanning laser Doppler vibrometer (see Figure 8(b); the shaker is at the

backside of the slat track, which is vertically suspended by means of a elastic rope). The FRF

measurements are performed in 500 points uniformly distributed over the surface of one side of the

slat track (see Figure 8(c)).

Starting from the FRF measurements, the modal parameters can be estimated using the frequency-

domain estimator given in Section 3. In the considered frequency band, a few hundreds of modes

are present. Figure 9 shows 4 of the estimated mode shapes (with enlarged displacement

amplitudes). The real displacements are of the order of a few micrometers.

(a) (b)

(c) (d)

Figure 9: Mode shapes of the slat track for a modal frequency of

(a) 4091 Hz, (b) 4161 Hz, (c) 4739 Hz, and (d) 5169 Hz

5. Conclusions

In this contribution about modal analysis, a ‘dedicated’ multivariable implementation for frequency-

domain estimators, based on a common-denominator transfer function model, has been given.

Typical for modal analysis is the very large number of outputs. This huge amount of data requires

dedicated algorithms that balance between accuracy and memory/computing requirements.

The results given in this contribution can be generalized to other frequency-domain estimators such

as the Total Least Squares (TLS), Generalized Total Least Squares (GTLS), Iterative Quadratic

Maximum Likelihood (IQML), … (see 6.43.8.2 Estimation with Known Noise Model and 6.43.8.4

Estimation with Unknown Noise Model).

Bibliography

Ewins D.J. (2000). Modal Testing: Theory, Practice and Application, Hertfordshire: Research

Studies Press. [In this book, all the steps involved in planning, executing, interpreting and applying

the results from a modal test are described in straightforward terms.]

Ewins D.J. and Inman D.J., Editors (2001). Structural Dynamics @ 2000: Current Status and

Future Directions, Baldock: Research Studies Press. [This book presents an integrated collection of

contributions on structural dynamics.]

Heylen W., Lammens S. and Sas P. (1998). Modal Analysis Theory and Testing, KULeuven (ISBN:

90-73802-61-X). [This book gives a good introduction to the theory as well as all practical aspects

of experimental modal analysis.]

Inman D.J. (1994). Engineering Vibration, Englewood Cliffs: Prentice Hall. [This book gives an

introduction to mechanical systems and vibration.]

Maia N.M.M. and Silva J.M.M., Editors (1997). Theoretical and Experimental Modal Analysis,

Taunton: Research Studies Press. [This book takes an advanced and up-to-date look at modal

analysis. The basics of vibration theory and signal processing are discussed.]

Pintelon R. and Schoukens J. (2001). System Identification: A Frequency Domain Approach, IEEE

Press and John Wiley & Sons (ISBN 0-7803-6000-1). [This book presents a general approach to

system identification, with both practical examples and theoretical discussions.]

1940. mechanical systems with a few inputs and hundreds of outputs have to be identified. curve-fit the data to estimate the resonance frequencies and damping ratios. the Tacoma Narrows suspension bridge collapsed due to wind-induced vibration (i. near the city of Tacoma. the bridge had only been open for traffic a few months. flight flutter tests have to be performed to detect possible onset of flutter. The damping values are then extrapolated in order to determine whether it is save to proceed to the next flight test point. 1. Figure 1: Tacoma Narrows Bridge Disaster. that small forces can result in important deformation. On November 7.Summary In this contribution the applicability of frequency-domain estimators in the field of modal analysis will be illustrated. Flutter will occur when one of the damping values tends to become negative. Introduction It is well known that (mechanical) structures can resonate. The Tacoma Narrows bridge disaster (Figure 1) is a typical example of this. Washington. A practical example will be given and finally the conclusions will be drawn. damage can be induced in the structure. i. The basics of vibration and modal analysis are briefly summarized. Before starting the flight tests. The classical flight flutter testing approach is to expand the flight envelope of a airplane by performing a vibration test at constant flight conditions. Wings of airplanes can be subjected to similar flutter phenomena during flight. . Situated on the Tacoma Narrows in Puget Sound. ground vibration tests as well as numerical simulations and wind tunnel tests (see Figure 2) are used to get some prior insight into the problem. and then to plot these frequencies and damping estimates against flight speed or Mach number. Before an airplane is released. flutter).e.e. In modal analysis. This requires adapted frequency-domain estimators designed to handle large amount of data in a reasonable amount of time. and possibly.

Today. its modes will change. In every development of a new or improved mechanical product.e. The majority of structures can be made to resonate. Resonant vibration is mainly caused by an interaction between the inertial and elastic properties of the materials within a structure. i. Figure 3: Modal analysis of a car body. the so-called “modal parameters”). . Resonance is often the cause of. or at least a contributing factor to many of the vibration and noise related problems that occur in structures and operating machinery. To better understand any structural vibration problem. For instance. and are determined by the material properties (mass. it will vibrate differently. modal damping. To understand this. and stiffness). if mass is added to a structure. to vibrate with excessive oscillatory motion.(a) (b) Figure 2: Wind tunnel tests on a scaled model of (a) a Cessna and (b) an Airbus A380. and boundary conditions of the structure. If either the material properties or the boundary conditions of a structure change.e. we will make use of the concept of single and multiple-degree-of-freedom systems. 2. Each mode is defined by a natural (modal or resonant) frequency. the resonant frequencies of a structure need to be identified and quantified. and a mode shape (i. damping. structural dynamics testing on product prototypes is used to assess its real dynamic behavior. modal analysis has become a widespread means of finding the modes of vibration of a machine or structure (Figure 3). The “Modal” Model Modes are inherent properties of a structure.

If. If c = 0 . a mass ∆m is added to the original mass m of the structure. are the poles of the system. equals the inverse of the dynamic stiffness H ( s) = 1 ms + cs + k 2 (4) Figure 4: SDOF system. − cx (t ) the (viscous) damping force. i. λ = ± iω n . the system is not damped and the poles becomes purely imaginary. and (5) ζ = c 2mω n = σ λ the damping ratio ( f d = f n 1 − ζ 2 ). the position of the mass when f (t ) ≡ 0 . X ( s ) = H ( s ) F ( s ) . − mx(t ) the inertial force. the damping coefficient c is usually very small resulting in a complex conjugate pole pair λ = −σ ± iω d with f d = ω d 2π the damped natural frequency.e.2. . This equation states that the sum of all forces acting on the mass m should be equal to zero with f (t ) an externally applied force. its natural frequency decreases to ω n = k ( m + ∆m ) . c the damping coefficient. i. for instance. In mechanical structures. The roots of the denominator of the transfer function. f n = ω n 2π the (undamped) natural frequency where ω n = k m = λ .1 Single Degree of Freedom A single-degree-of-freedom (SDOF) system (see Figure 4 where the mass m can only move along the vertical x-axis) is described by the following equation mx (t ) + cx (t ) + kx (t ) = f (t ) (1) with m the mass. d ( s ) = ms 2 + cs + k . and − kx (t ) the restoring force. Transforming (1) to the Laplace domain (assuming zero initial conditions) yields Z ( s) X ( s) = F ( s) (2) with Z (s ) the dynamic stiffness Z ( s ) = ms 2 + cs + k (3) The transfer function H (s ) between displacement and force.e. and k the stiffness. The variable x (t ) stands for the position of the mass m with respect to its equilibrium point.

d ( s ) = det( Ms 2 + Cs + K ) (12) (11) . Although very few practical structures could realistically be modeled by a single-degree-of-freedom (SDOF) system. f2(t) 0 m M= 1 f (t ) 0 m2 m2 f (t ) = 1 x2(t) f1(t) f 2 (t ) c2 k1 + k2 − k2 k2 K= k2 m1 − k2 x (t ) x1(t) x(t ) = 1 c1 x2 (t ) k1 c1 + c2 − c2 C= c2 − c2 Figure 5: 2-DOF system. Transforming (7) to the Laplace domain (assuming zero initial conditions) yields Z( s )X ( s ) = F( s ) (8) with Z(s ) the dynamic stiffness matrix Z ( s ) = Ms 2 + C s + K (9) The transfer function matrix H(s ) between displacement and force vectors. denoted by H (ω ) . then H (ω ) goes to infinity for ω → ω n = k m (see Figure 4).2 Multiple Degree of Freedom Multiple-degree-of-freedom (MDOF) systems are described by the following equation Mx (t ) + Cx (t ) + Kx(t ) = f (t ) (7) In Figure 5. also known as the characteristic polynomial. X ( s ) = H( s )F( s ) .The Frequency Response Function (FRF). is obtain by replacing the Laplace variable s in (4) by iω resulting in H (ω ) = 1 1 = − mω + icω + k ( k − mω 2 ) + icω 2 (6) Clearly. 2. equals the inverse of the dynamic stiffness matrix H( s ) = [Ms 2 + Cs + K ]−1 = N( s ) d ( s) (10) with the numerator polynomial matrix N(s ) given by N( s ) = adj( Ms 2 + Cs + K ) and the common-denominator polynomial d (s ) . the properties of such a system are important because those of a more complex multiple-degree-of-freedom (MDOF) system can always be represented as the linear superposition of a number of SDOF characteristics (when the system is linear time-invariant). the different matrices are defined for a 2-DOF system with both DOF along the vertical x-axis. if c = 0 .

there will always be a (small) contribution of the other modes resulting in X (ω d . From equation (13). m = 1. with N m the number of modes of the system. When σ n is small. m = 1. Applying a harmonic force at one of the DOFs. results in a displacement vector that is approximately equal to X (ω d . N m .n ) = n N ∗ ∗ ψ n (r) N mψ m ( r ) mψ m ( r ) +∑ +∑ Fr (ω d .e. i. say r. the roots of the characteristic polynomial d (s ) are complex conjugate pole pairs. s = iω d . In reality. the so-called “modal” model (assuming all poles have multiplicity one) Rm R∗ m H( s ) = ∑ + s − λ∗ m =1 s − λm m Nm (13) The residue matrices R m . The transfer m function can be rewritten in a pole-residue form.n is called an “operating deflection shapes” and is approximately proportional to the mode shape vector of mode n. i. . . the proportionality is well satisfied. the poles λm = −σ m ± iω d .n ) σn (17) The observed displacement at angular frequency ω d . i. the overall vibration shape (“operating deflection shape”) of a structure will tend to be dominated by the mode shape of the resonance.2. n . . are defined by R m = lim H( s )( s − λm ) s →λm (14) It can be shown that the matrix R m is of rank one meaning that R m can be decomposed as Rm = m T m ψ m (1) ψ ( 2) = m ψ m (1) ψ m ( 2) ψ m ( N m ) ψ m ( N m ) (15) with m a vector representing the “mode shape” of mode m.n ) σn m =1 σ m + i (ω n − ω m ) m =1 σ m + i(ω n + ω m ) m ≠n m m (18) . one concludes that the transfer function matrix of a linear time-invariant MDOF system with N m DOFs is the sum of N m SDOF transfer functions (“modal superposition”) and that the full transfer function matrix is completely characterized by the modal parameters.n ) ≈ n ψ n (r) Fr (ω d .1 Mode Shapes and Operating Deflection Shapes At or near the natural frequency of a mode. i. m = 1.e.n . 2. N m . with angular frequency corresponding to the damped natural frequency of for instance the n-th mode. Taking the inverse Laplace transform of (13) gives the Impulse Response Function (IRF) h(t ) = ∑ R m e λmt + R ∗ e λmt m ∗ Nm (16) m =1 which consists of a sum of complex exponential functions.e.When the damping is small.m and the mode shape vectors m . λm and λ∗ . N m .e.

. Based on these results.43. mode n cannot be excited (or “controlled”) by applying a force in DOF 1. The same is true for the response measurements. it is possible to implement more sophisticated identification methods (see 6.2.1. In practice. Several choices are possible for the polynomial basis functions Ω j (ω ) .1 Least Squares Estimation 3.e. for example.1 ( s ) = ∑ ∗ ∗ ψ m ( 2) ⋅ψ m (1) ψ m ( 2) ⋅ψ m (1) + s − λm s − λ∗ m =1 m Nm (19) If ψ n (1) ≠ 0 then mode n will only be “observed” in DOF 2 if ψ n (2) ≠ 0 .8.2. the multiple-input-multiple-output (MIMO) transfer function matrix simplifies to the following single-input-single-output (SISO) transfer function H 2. This huge amount of data requires dedicated algorithms that balance between accuracy and memory/computing needs.. while the number of outputs (response measurements) can reach more than 1000 points when using optical measurement equipment such as for instance a scanning laser Doppler vibrometer.2 Estimation with Known Noise Model and 6. 3.4 Estimation with Unknown Noise Model) such as for instance the frequency-domain Maximum Likelihood (ML) estimator (Section 3. the number of excitation points can be increased.8. this means that the force actuator should not be positioned in a nodal point of the modes of interest. The number of inputs (excitation points) is typically in the order of 1 to 10. To reduce the risk of missing modes. The bad .1 Common-Denominator Model The relationship between output o ( o = 1. In that case.2). the functions Ω j (ω ) are usually given by Ω j (ω ) = exp( −iωTs ⋅ j ) (with Ts the sampling period) while for a continuous-time domain model Ω j (ω ) = (iω ) j ..43. The DOFs where a mode shape vector equals zero are called nodal points or nodes.1 such a ‘dedicated’ frequency-domain least-squares estimator will be presented. If ψ n (1) = 0 then it is clear that the terms corresponding to mode n will not appear in the sum. Frequency-Domain Identification of Modes Typical for modal analysis is the very large number of outputs. N o ) and input i ( i = 1. For a discrete-time domain model.2 Observability and Controllability of Modes Assuming. i. that one force is applied in DOF 1 while the displacement is observed in DOF 2. The real-valued coefficients A j and Bkj are the parameters to be estimated. In Section 3. N i ) is modeled in the frequency domain by means of a common-denominator transfer function N (ω ) ˆ H k (ω ) = k d (ω ) for k = 1. 3. N o N i (where k = (o − 1) N i + i ) and with (20) N k (ω ) = ∑ Ω j (ω ) Bkj j =0 n (21) the numerator polynomial between output o and input i and d (ω ) = ∑ Ω j (ω ) A j j =0 n (22) the common-denominator polynomial.

Ωn (ω1 )] Xk = Wk (ω N )[ Ω 0 (ω N )..43. Ωn (ω N f f f f (26) − Wk ( ω1 )[ Ω 0 ( ω1 ) H k (ω1 ). Replacing the model H k (ω ) in (20) by the measured FRFs H k (ω f ) for f = 1. Ω1 (ω N f )H k (ω N f ). 3. . Ωn (ω1 )H k (ω1 )] Yk = − Wk (ω N f )[ Ω 0 (ω N f )H k (ω N f ). N o N i and f = 1. Ω1 (ω N ). Because the equations (23) are “linear-in-the-parameters”. Ω1 (ω1 )H k (ω1 ). . The (complex) Jacobian matrix J of this least-squares problem X1 0 J= 0 0 0 X2 0 0 X No Ni Y1 Y2 YN o N i (28) has Nf No Ni rows and (n+1)(No Ni +1) columns (with Nf >> n. . measurements of Frequency Response Functions (FRFs) are commonly used (see ˆ 6. N f gives. ∑Ω j =0 n j (ω f )Bkj − ∑ Ω j (ω f )H k (ω f ) A j ≈ 0 j =0 n (23) with k = 1. they can be reformulated as X 1 0 0 0 X2 0 X No Ni 0 0 Y1 1 Y2 2 ⋅ ≈ 0 YN o N i N o N i (24) with k Bk 0 B = k1 . Ω1 (ω1 ). The quality of the estimates can often be improved by using an adequate weighting function. .. N f . .2 Linearity in the Parameters In modal analysis.. where n is the order of the polynomials.8.numerical conditioning of the continuous-time domain approach can be improved by using for instance orthogonal Forsythe polynomials (at the expense of an increase of the computation time). Bkn A0 A = 1 An )] )] (25) Wk ( ω1 )[ Ω 0 (ω1 ). n = 2 N m ). after multiplication with the denominator polynomial.1. Ωn (ω N f )H k (ω N f (27) Note that every equation in (23) has been weighted with a frequency-dependent function Wk ( ω f ) .1 Measurements of Frequency response functions).

As we are mainly interested in a fast and stable method to construct a stabilization chart (see next section).4 Reduced Normal Equations Although the number of rows of the normal matrix in (29) is much smaller than the number of rows of the Jacobian matrix (28).e.e.e. ω f = f ⋅ ∆ω . Elimination of the numerator coefficients k = − R −1 ⋅ S k ⋅ k (32) yields No Ni T −1 ∑ Tk − S k ⋅ R k ⋅ S k ⋅ ≈ 0 k =1 (33) .3 Normal Equations Many estimators used in modal analysis form the normal equations explicitly. i. Note that the real part of J H J has to be taken because the coefficients are real. s ) = Re ∑ Wk (ω f ) ⋅ e i2π ( r − s ) f f =1 N N Nf 2 S k ( r. s ) = − Re ∑ Wk (ω f ) H k (ω f ) ⋅ Ω r −1 (ω f )Ω s −1 (ω f ) f =1 Nf 2 H Tk ( r. s ) = Re ∑ Wk (ω f ) H k (ω f ) ⋅ Ω r −1 (ω f )Ω s −1 (ω f ) f =1 (30) If a discrete time-domain model is used. s ) = Re f =1 N (31) One can readily verify that the above matrices have a Toeplitz structure and that their entries can be time-efficiently computed with the Fast Fourier Transform (FFT) algorithm. the poles) are in fact required. Ω j (ω f ) = exp( −iω f Ts ⋅ j ) .e. they compute Re( J H J ) explicitly. as will be the case here too. f = 1. and if the frequencies are uniformly distributed (i. However. its size is still quite huge (i.e. The normal equations can be written as R 1 0 T S1 0 R2 ST 2 1 2 ⋅ ≈ 0 No Ni ∑ Tk N o N i k =1 S1 S2 (29) H H with R k = Re( X k X k ) . with ∆ω = 2π NTs ). N f . s ) = Re ∑ Wk (ω f ) ⋅ Ω r −1 (ω f )Ω s −1 (ω f ) f =1 Nf 2 H S k ( r. i. forming the normal equations can result in a faster implementation. 3.1. S k = Re( X k Yk ) . then.3..1. Deriving the estimates directly from the Jacobian matrix leads to a better-conditioned problem. and Tk = Re( YkH Yk ) . (n+1)(No Ni +1) rows and columns). only the denominator coefficients (i. s ) = − Re ∑ Wk (ω f ) H k (ω f ) ⋅ e i2π ( r − s ) f f =1 Nf 2 ∑ Wk (ω f ) H k (ω f ) ⋅ e i2π ( r − s ) f Tk ( r. The entries of these matrices equal Nf 2 H Rk ( r. the above summations can be rewritten as Nf 2 Rk ( r.

In Figure 6 the stabilization chart of the LSCE estimator is compared with the proposed frequency-domain least squares estimator. A stabilization chart is obtained by repeating the analysis for increasing model order n. the poles with a negative real part) are then presented graphically in ascending model order in a so-called “stabilization chart” (see Figure 6).1. To remove the parameter redundancy of transfer function model (20) (and to avoid the trivial solution with all coefficient equal to zero). These mathematical poles are mainly due to the presence of noise on the measurements. It turns out that in many applications. tend to jump around. for instance.e. for instance. (32) can be used to derive all ˆ LS coefficients. Stabilization chart obtained with (a) a time-domain estimator (LSCE) and (b) the frequency-domain least-squares estimator. For each model order. The LSCE estimator is a time-domain technique that makes use of impulse response functions (16) to derive the modal parameters.1 : n )]−1 ⋅ {M(1 : n. The mode shape vectors are derived using (14) and (15).5 Stabilization chart In modal analysis. and thus k k much smaller than the original normal equation (29). In that case.o or M ⋅ ≈ 0 with M = ∑k =1 i Tk − S T ⋅ R −1 ⋅ S k . n + 1)} ˆ LS = 1 (34) Once ˆ LS is known. the poles are calculated from the estimated denominator coefficients. while the so-called mathematical poles. the Least Squares (LS) estimate of D is given by N N − [M(1 : n. that the last coefficient of D is constrained to 1 (i. the frequency-domain estimator is able to generate quite clear stabilization compared to the LSCE approach. The size of the square matrix M is n+1. Assume. i. The LSCE estimator (Least Squares Complex Exponential) is probably the most frequently used technique in industry. a constraint has to be imposed on the coefficients of the transfer functions. (a) (b) Figure 6. 3. Estimated poles corresponding to physically relevant system modes tend to appear for each estimation order at nearly identical locations.e. This approach is more time efficient that solving (29) directly (approximately N o2 N i2 times faster). . by imposing that one of the coefficients is equal to a non-zero constant value. coefficient n+1). a stabilization chart is an important tool that is often used to assist the user in separating physical poles from mathematical ones. This can be done. poles resulting from the mathematical solution of the normal equations but meaningless with respect to the physical interpretation.e. The stable poles (i.

From (39). As one is mainly interested in the uncertainty on the modal frequencies and damping ratios. Re( J H J p ) and Re( J H rp ) ) in a similar time-efficient way as p p presented in Section 3.2 Estimation with Known Noise Model and 6. T ]T is obtained by minimizing (35). .43.1 Gauss-Newton Optimization Assuming the FRFs to be uncorrelated. which takes advantage of the quadratic form of the cost function (35). one can show that this matrix is given by No Ni − cov{ ˆ ML } ≈ 2 ∑ Tk − S T ⋅ R k 1 ⋅ S k k k =1 −1 (39) H with R k . . ω f ) − H k (ω f ) var{H k (ω f )} 2 (35) T The Maximum Likelihood (ML) estimate of = [ 1 . the (negative) log-likelihood function reduces to " ML ( ) = No Ni N f ∑∑ k =1 f =1 ˆ H k ( . The Gauss-Newton iterations are given by (a) solve Re( J H J p ) p (b) set p +1 p p = − Re( J H rp ) for p p = p + (36) with rp = r( p ) .2 Maximum Likelihood Estimation 3. For flight flutter testing – but also for applications such as vibration-based fault detection and operational modal analysis – the availability of reliable estimates together with confidence intervals is important. ω N f ) − H N o N i (ω N f ) var{H N o N i (ω N f )} (37) The Jacobian matrix J p has the same structure as the matrix J given in (28).8.3.8. Also here it is possible to form the normal equations (i.43. 3.43. it is not necessary to invert the full matrix occurring in (38).2. J p = ∂r ( ) ∂ and p ˆ H 11 ( . ω1 ) − H 11 (ω1 ) var{H 11 (ω1 )} r( ) = ˆ H N o N i ( .2 Confidence Intervals A good approximation of the covariance matrix of the ML estimate ˆ ML is obtained by inverting the Fisher information matrix (see 6.4 Estimation with Unknown Noise Model for more information about frequency-domain ML identification.8. Hence.2. Frequency Domain System Identification) H cov{ ˆ ML } ≈ [2 Re( J ∞ J ∞ )]−1 (38) with J ∞ the Jacobian matrix evaluated in the last iteration step of the Gauss-Newton optimization. it is possible to compute the uncertainty on the modal frequencies and damping ratios. S k .3 but now applied to Re( J ∞ J ∞ ) . and Tk as defined in Section 3. See 6. N This can be done by means of a Gauss-Newton optimization algorithm. only the covariance matrix of the denominator coefficients is in fact required. Starting from (38). T o N i .1.1.e.

It is commonly accepted that the track should outlive the plane by five times. (a) Excitation with a shaker. Safety critical components such as slat tracks are rigorously tested to prove their ability to withstand all safety regulations. The A320 airplane has 5 slats per wing. SLAT TRACK LEADING EDGE OF WING SLAT Figure 7: Mounted slat track on an Airbus A320 in extended position. Application In this section. modal analysis will be applied to a slat track. To validate the dynamic behavior of these FE models. The first slat (i. it is possible to predict the lifetime of a track using Finite Element (FE) models. (b) Response measurements with a scanning laser Doppler vibrometer. A slat track of an Airbus A320 airplane is considered here. experimentally obtained estimates of the modal parameters are required. A force sensor is attached on the slat track and is in contact with the shaker through a stinger. The stinger has the characteristic of being stiff in only one direction. that of the intended excitation. which is safety critical component of a airplane.4. Slat tracks (see Figure 7) are located at the leading edge of an aircraft wing and make part of a gliding mechanism that is used to enlarge the wing surface. In this example. The other 4 slats have 2 slat tracks each. A multisine is a periodic signal consisting of a sum of sine waves at uniformly distributed frequencies (from 0 Hz to 8192 Hz in this case with a resolution of 1 Hz). In Figure 8(a) a shaker is used to excite the slat track with a multisine signal. Using computer simulations. the inboard slat between fuselage and engine) contains 4 slat tracks.e. one surface of the track is . SLAT TRACK FORCE SENSOR STINGER SHAKER (a) SHAKER SCANNING LASER DOPPLER VIBROMETER (b) (c) Figure 8: Measurement setup. i.e. (c) Measurement points on the surface of the slat track. The enlargement of the wing surface is needed in order to increase the lift force at reduced velocity during landing and take off.

(a) (b) (c) (d) Figure 9: Mode shapes of the slat track for a modal frequency of (a) 4091 Hz. which is vertically suspended by means of a elastic rope). In the considered frequency band. the shaker is at the backside of the slat track. Starting from the FRF measurements. and (d) 5169 Hz . a few hundreds of modes are present.measured by means of a scanning laser Doppler vibrometer (see Figure 8(b). The FRF measurements are performed in 500 points uniformly distributed over the surface of one side of the slat track (see Figure 8(c)). (c) 4739 Hz. (b) 4161 Hz. The real displacements are of the order of a few micrometers. Figure 9 shows 4 of the estimated mode shapes (with enlarged displacement amplitudes). the modal parameters can be estimated using the frequencydomain estimator given in Section 3.

Englewood Cliffs: Prentice Hall. [This book presents an integrated collection of contributions on structural dynamics.M. (1998).] Inman D. System Identification: A Frequency Domain Approach. executing. has been given. Bibliography Ewins D. interpreting and applying the results from a modal test are described in straightforward terms.M.J. (2000).4 Estimation with Unknown Noise Model). with both practical examples and theoretical discussions.M. [This book gives an introduction to mechanical systems and vibration.J. Editors (1997). The results given in this contribution can be generalized to other frequency-domain estimators such as the Total Least Squares (TLS). all the steps involved in planning.. Conclusions In this contribution about modal analysis. Lammens S. Typical for modal analysis is the very large number of outputs.8.43. Theoretical and Experimental Modal Analysis. [This book presents a general approach to system identification.. The basics of vibration theory and signal processing are discussed.] Maia N..2 Estimation with Known Noise Model and 6. [In this book. This huge amount of data requires dedicated algorithms that balance between accuracy and memory/computing requirements. Iterative Quadratic Maximum Likelihood (IQML). … (see 6. KULeuven (ISBN: 90-73802-61-X). and Sas P. and Schoukens J.] Ewins D. Modal Testing: Theory. (1994). a ‘dedicated’ multivariable implementation for frequencydomain estimators.43. Structural Dynamics @ 2000: Current Status and Future Directions. and Silva J. Modal Analysis Theory and Testing. Editors (2001).] .8. IEEE Press and John Wiley & Sons (ISBN 0-7803-6000-1). Hertfordshire: Research Studies Press. Engineering Vibration. Generalized Total Least Squares (GTLS). (2001).J. [This book gives a good introduction to the theory as well as all practical aspects of experimental modal analysis.5. and Inman D. Taunton: Research Studies Press. based on a common-denominator transfer function model. Baldock: Research Studies Press. [This book takes an advanced and up-to-date look at modal analysis.] Pintelon R.] Heylen W.J.M. Practice and Application.

Sign up to vote on this title

UsefulNot useful- Modal Analysis
- 06_Experimental Modal Analysis
- Theoretical and Experimental Modal Analysis
- Modal Analysis
- modalBK
- Modal Analysis
- Experimental Modal Analysis
- shaker excitation for experimental modal analysis
- ANSYS Structural Analysis Guide MODAL ANALYSIS
- B&K Structural Testing
- Modal Analysis
- Modal Analysis
- Modal Analysis
- Cyclic Symmetric Modal Analysis
- Fundamentals Modal Testing
- Fundamentals of Vibration and Modal Analysis
- 02-Modal Testing and Analysis
- Adams Help
- Modal Testing_Theory and Practice 2nd Edition
- 20224644 Vibration Fundamentals
- D. J. Ewins - Modal Testing
- MSC.ADAMS TUTORIAL
- Vibration
- Ewins D. J., Modal Testing - Theory and Practice
- Vibration of pumps
- M. Rades - Mechanical Vibrations 1
- Blevins_Formulas for Natural Frequency and Mode Shape
- Vibrations of Shells and Plates
- Saeid Diss
- modal analysis